Practical terminology map · 41 visible definitions

Motion intelligence glossary.

The motion intelligence glossary connects the language of AI animation, digital humans, robotics, humanoids, control and actuation. Each definition identifies what a term means, how it is used and where it fits in the wider motion stack.

Conceptual research resource · Not an operating product or research institution

These definitions provide a practical conceptual map across several evolving technical fields. Terminology may differ across research groups, companies and implementations.

Alphabetical index

Find a term directly.

  1. Actuation
  2. Actuator
  3. AI motion generation
  4. Avatar behaviour
  5. Dexterous manipulation
  6. Digital human
  7. Embodied AI
  8. Embodied reasoning
  9. Feedback control
  10. Forward kinematics
  11. Generative motion
  12. Human motion synthesis
  13. Imitation learning
  14. Impedance control
  15. Inverse kinematics
  16. Locomotion
  17. Manipulation
  18. Motion capture
  19. Motion coherence
  20. Motion control
  21. Motion diffusion model
  22. Motion intelligence
  23. Motion planning
  24. Motion retargeting
  25. Motion style transfer
  26. Physical AI
  27. Reinforcement learning
  28. Robot policy
  29. Robot policy learning
  30. Robotics simulation
  31. Sim-to-real
  32. Synthetic data
  33. Temporal consistency
  34. Text-to-motion
  35. Torque control
  36. Trajectory
  37. Trajectory generation
  38. Vision-language-action model
  39. Whole-body control
  40. World foundation model
  41. World model

13 terms

Generative motion

Terms used to represent, synthesise, edit and evaluate movement in digital characters, avatars and visual sequences.

AI motion generation

AI motion generation uses computational models to create movement over time from signals such as text, images, audio, video, action labels, key poses or demonstrations.

The output may be an explicit pose or trajectory sequence, an editable animation representation, or pixels in a generated video. Quality depends on semantic alignment, temporal coherence, controllability and the target environment.

Generation layer · digital motion

Place ai motion generation in its pillar context →

Avatar behaviour

Avatar behaviour is the observable pattern of gesture, posture, gaze, timing and response through which a digital representation acts in an environment.

It extends beyond appearance. Behaviour connects motion generation with interaction logic, user input, social signalling and continuity across an experience.

Execution and interaction · digital motion

Place avatar behaviour in its pillar context →

Digital human

A digital human is a computationally represented human-like character whose appearance, motion, voice or behaviour can be rendered and controlled.

Digital humans range from authored characters to interactive agents. Human likeness does not by itself imply autonomy; motion and behavioural systems determine how the representation acts over time.

Character and interaction layer

Place digital human in its pillar context →

Generative motion

Generative motion is movement produced by a model as one or more plausible sequences rather than retrieved as a fixed recorded clip.

The model may condition on intent, context, style, geometry or partial motion. Generative output can be diverse, but useful systems also need editability and constraints.

Generation layer

Place generative motion in its pillar context →

Human motion synthesis

Human motion synthesis is the computational creation of pose sequences that depict human movement.

Methods can be procedural, example-based, optimisation-based or learned. Modern systems often condition on language, action, audio, scene context or reference motion.

Generation layer · articulated motion

Place human motion synthesis in its pillar context →

Motion capture

Motion capture records the movement of a person or object and converts it into data that can drive or inform a digital representation.

It supplies measured examples rather than generating movement from scratch. Captured motion may still need cleaning, segmentation, retargeting and physical interpretation.

Data and representation layer

Place motion capture in its pillar context →

Motion coherence

Motion coherence is the degree to which movement remains internally consistent across time, body parts, contacts and intended action.

A coherent sequence avoids unexplained jumps, disconnected limb timing and contradictory trajectories. Coherence can be visual, kinematic, semantic or physical.

Temporal refinement and evaluation

Place motion coherence in its pillar context →

Motion diffusion model

A motion diffusion model generates a motion sequence by learning to reverse a process that progressively corrupts motion data with noise.

Conditioning signals such as language or keyframes can guide denoising toward plausible movements. Diffusion supports diversity but does not automatically guarantee contact quality or physical execution.

Generative-model layer

Place motion diffusion model in its pillar context →

Motion intelligence

Motion intelligence is the capability to generate, interpret, plan, adapt and control movement across time.

Animatio.ai uses the term as a practical bridge across established digital and physical motion fields, not as a claim that one universal scientific taxonomy already exists.

Connecting thesis across the complete stack

Place motion intelligence in its pillar context →

Motion retargeting

Motion retargeting transfers movement from one skeleton, character or embodiment to another with different proportions or joint structure.

A useful retarget preserves task-relevant timing and relationships while respecting target joint limits, contacts and style. For robots, dynamic feasibility adds another constraint layer.

Representation and adaptation

Place motion retargeting in its pillar context →

Motion style transfer

Motion style transfer changes qualities such as rhythm, energy, posture or expression while preserving a movement’s underlying action.

Separating content from style is imperfect because movement qualities can alter timing and contact. Control should therefore be assessed across the full sequence.

Generation and refinement

Place motion style transfer in its pillar context →

Temporal consistency

Temporal consistency is the preservation of stable appearance, identity, geometry and motion relationships from one moment to the next.

In generated video it includes object and character continuity. In explicit motion it includes smooth pose evolution, velocity, contact and action phase.

Temporal refinement and evaluation

Place temporal consistency in its pillar context →

Text-to-motion

Text-to-motion is the generation of a movement sequence from a natural-language description.

The text may name an action, describe manner and timing, or specify interaction. Ambiguous prompts can admit multiple valid motions, so diversity and controllability both matter.

Conditioning and generation

Place text-to-motion in its pillar context →

5 terms

Kinematics and trajectories

Terms describing articulated geometry, paths through space and the conversion of goals into feasible movement.

Forward kinematics

Forward kinematics calculates the position and orientation of an articulated system’s links from known joint values.

It answers where the end effector or body parts will be for a given configuration. It does not choose the joint values or test dynamics by itself.

Representation and geometry

Place forward kinematics in its pillar context →

Inverse kinematics

Inverse kinematics solves for joint values that place an end effector or body part at a desired pose.

There may be multiple solutions, no exact solution or solutions that violate collision, balance or actuator limits. Constraints and optimisation determine which answer is useful.

Planning and geometric control

Place inverse kinematics in its pillar context →

Motion planning

Motion planning searches for a path or trajectory that connects states while respecting obstacles and specified constraints.

A planner may work in joint space or task space and can use sampling, search or optimisation. Planning is distinct from the controller that tracks the result.

Planning layer

Place motion planning in its pillar context →

Trajectory

A trajectory specifies how a system’s state should evolve over time, often including positions, velocities and accelerations.

Unlike a purely geometric path, a trajectory has timing. That timing affects smoothness, forces, feasibility and controller demands.

Planning output and control reference

Place trajectory in its pillar context →

Trajectory generation

Trajectory generation creates a time-parameterised sequence that carries a system from an initial state toward a goal.

It can interpolate waypoints, optimise an objective or translate a learned action into a continuous reference while enforcing velocity and acceleration limits.

Planning-to-control interface

Place trajectory generation in its pillar context →

13 terms

Robot intelligence and learning

Terms connecting perception, reasoning, learned action, predictive models, simulation and transfer.

Embodied AI

Embodied AI studies intelligent systems whose perception and action are shaped by an environment and a body, physical or simulated.

The embodiment determines available observations, actions and constraints. Intelligence is therefore evaluated through interaction, not only static prediction.

System-level field

Place embodied ai in its pillar context →

Embodied reasoning

Embodied reasoning connects semantic understanding with spatial, physical and action consequences in an environment.

A system must reason not only about what an instruction means but about reachability, geometry, contact, sequence and the effects of acting through its particular body.

Reasoning and task-planning layer

Place embodied reasoning in its pillar context →

Imitation learning

Imitation learning trains a policy to reproduce behaviour demonstrated by an expert, recorded trajectory or reference motion.

It can reduce manual reward design but inherits limits from the demonstrations and the mismatch between demonstrator and target embodiment.

Policy-learning method

Place imitation learning in its pillar context →

Physical AI

Physical AI is a broad contemporary label for AI systems that perceive, reason and act in the physical world.

Usage varies across organisations. The label can include robotics, autonomous systems, simulation, world models and embodied learning, but it is not one standardised architecture.

Application territory across the physical stack

Place physical ai in its pillar context →

Reinforcement learning

Reinforcement learning trains an agent to select actions by optimising expected reward through interaction with an environment.

For motion, rewards may encode task success, energy, style, stability or safety. Simulation is often used because physical exploration can be slow or unsafe.

Policy-learning method

Place reinforcement learning in its pillar context →

Robot policy

A robot policy is a rule or learned function that maps observations or states to actions.

Actions can be high-level skills, target poses, trajectories, joint commands or actuator-level controls. The action space changes what the policy must learn.

Policy layer

Place robot policy in its pillar context →

Robot policy learning

Robot policy learning uses data or interaction to acquire a mapping from observations and goals to robot actions.

Common sources include demonstrations, reinforcement, offline trajectories and cross-robot datasets. Learned policies still operate within hardware and safety constraints.

Learning layer

Place robot policy learning in its pillar context →

Robotics simulation

Robotics simulation models a robot, its environment and their physical interaction in software.

It supports planning tests, policy training, synthetic data and failure analysis. Accuracy depends on geometry, contact, sensors, actuators and system identification.

Simulation and validation layer

Place robotics simulation in its pillar context →

Sim-to-real

Sim-to-real is the transfer of a model, policy or behaviour developed in simulation to a physical system.

The reality gap can involve visual appearance, contact, latency, friction, actuator response and sensor noise. Domain randomisation, calibration and real-world adaptation help reduce it.

Transfer and deployment

Place sim-to-real in its pillar context →

Synthetic data

Synthetic data is information generated by simulation, procedural systems or learned models rather than captured directly from the target real-world process.

It can expand coverage and label availability, but utility depends on whether generated variation preserves the factors that matter for the downstream task.

Data layer

Place synthetic data in its pillar context →

Vision-language-action model

A vision-language-action model connects visual observations and language instructions with actions for an embodied system.

Some models tokenise actions; others use specialised continuous action heads. The category seeks semantic generalisation while meeting embodiment and control demands.

Perception, reasoning and policy

Place vision-language-action model in its pillar context →

World foundation model

A world foundation model is a broadly pretrained model intended to support prediction or generation of environment dynamics across multiple downstream physical-AI tasks.

The term is evolving and vendor usage differs. Scale does not remove the need for action grounding, physical evaluation and domain-specific adaptation.

Predictive-model and data layer

Place world foundation model in its pillar context →

World model

A world model predicts relevant aspects of how an environment or system state may change after actions or events.

Predictions can occur in pixels, explicit state or a learned latent space. A controller or planner can use them to compare possible futures before acting.

Prediction and planning support

Place world model in its pillar context →

6 terms

Physical execution and control

Terms describing how planned or learned movement becomes continuously corrected physical action.

Actuation

Actuation is the conversion of a command or energy input into physical movement or force.

It is the execution layer of a machine. Actuation quality affects bandwidth, precision, compliance and force, but does not alone supply planning or intelligence.

Physical execution layer

Place actuation in its pillar context →

Actuator

An actuator is a component that produces controlled motion or force, commonly using electric, hydraulic or pneumatic power.

Motors, drives, transmissions and sensors may be integrated differently. A controller sends targets within the actuator’s physical and thermal limits.

Physical execution component

Place actuator in its pillar context →

Feedback control

Feedback control adjusts commands using measured differences between desired and observed behaviour.

The loop can correct disturbance and modelling error, but its performance depends on sensing, latency, controller design and actuator response.

Control and adaptation layer

Place feedback control in its pillar context →

Impedance control

Impedance control regulates the relationship between motion and force so a robot exhibits a desired mechanical response to interaction.

Rather than demanding rigid position tracking, it can make contact behaviour more compliant. Tuning must reflect the task, controller rate and hardware.

Interaction-control layer

Place impedance control in its pillar context →

Motion control

Motion control converts desired movement into commands that continuously guide a system’s joints, motors or actuators.

It uses feedback to track trajectories, regulate force or maintain stability. Control operates closer to execution than high-level generation or planning.

Control layer

Place motion control in its pillar context →

Torque control

Torque control commands or regulates rotational force at a joint or motor.

It can support dynamic movement and compliant contact but requires accurate sensing, modelling and safety limits. Torque targets are not high-level task plans.

Low-level control

Place torque control in its pillar context →

4 terms

Humanoid skill

Terms describing locomotion, manipulation and coordinated control across complex articulated bodies.

Dexterous manipulation

Dexterous manipulation controls multiple contacts and degrees of freedom to reposition, use or transform objects with precision.

It may involve multi-fingered hands, tools or coordinated arms. Contact uncertainty, friction and sensing make the problem substantially harder than simple reaching.

Task execution and contact control

Place dexterous manipulation in its pillar context →

Locomotion

Locomotion is self-propelled movement of an embodied system from one place to another.

For legged robots it requires contact sequencing, balance, terrain adaptation and recovery. A visually plausible gait is not necessarily dynamically stable.

Whole-body planning and execution

Place locomotion in its pillar context →

Manipulation

Manipulation is the controlled physical interaction with objects to move, reorient, assemble or otherwise change them.

It combines perception, grasp or contact planning, trajectory generation and force-aware control. Whole-body systems may use posture and locomotion to extend reach.

Task planning, motion and control

Place manipulation in its pillar context →

Whole-body control

Whole-body control coordinates many joints, limbs and contacts around shared task and stability objectives.

It resolves interactions among posture, balance, locomotion and manipulation while considering kinematic, dynamic and actuator constraints.

Planning and control across the body

Place whole-body control in its pillar context →